DiskANN
DiskANN is a disk-oriented approximate-nearest-neighbor indexing system and a common baseline for papers that trade recall, latency, update cost, and SSD I/O behavior in large-scale vector search.
是什么
DiskANN-style designs keep a graph index with data laid out to make SSD-assisted traversal practical. In this wiki it is most often a baseline rather than a universal optimum: static search, dynamic insertion, memory footprint, and hardware placement stress different parts of the design.
关键观察 / 隐含假设
- 观察：update behavior changes the relevant baseline. OdinANN-FAST26 contrasts direct insertion and merge-related costs with disk-oriented graph indexing.
- 观察：memory and storage placement can alter ANN trade-offs. LEANN-MLSys26 and PIMANN-ATC25 study different resource placements around ANN search.
- 假设：a recall/latency comparison captures system utility. Terminus-MLSys26 illustrates that workload, update rate, and device behavior can introduce additional boundaries.
演进时间线
- 2025 ATC：PIMANN-ATC25 — explores an alternative hardware placement for ANN operations.
- 2026 FAST：OdinANN-FAST26 — addresses update-oriented ANN indexing trade-offs.
- 2026 MLSys：LEANN-MLSys26 — treats memory footprint and retrieval behavior as coupled constraints.
相关概念
相关论文
- OdinANN-FAST26 — update-oriented comparison with graph indexing.
- LEANN-MLSys26 — memory-efficient ANN retrieval.
- PIMANN-ATC25 — hardware placement for ANN search.
- Terminus-MLSys26 — system boundaries in ANN execution.